The researchers at Xiamen University propose a Sequential Action induced invariant Representation (SAR) method to accurately learn task-relevant state representations from high-dimensional observations with visual distractions. This is a challenging aspect of visual reinforcement learning. The SAR method optimizes the encoder to only preserve components that follow the control signals of sequential actions, allowing the agent to learn robust representations against distractions. The method is tested on DeepMind Control suite tasks and also deployed to real-world CARLA-based autonomous driving with natural distractions. The code and demo videos are available on the project page.
Publication date: 22 Sep 2023
Project Page: https://github.com/DMU-XMU/SAR.git
Paper: https://arxiv.org/pdf/2309.12628
